Choisir la langue :

Beyond the One Step Greedy Operator

Institutional tag: 
Thematic tag(s): 

I will describe a variation of dynamic programming for solving
infinite-horizon Markov Decision Process problems, where one puts more
computational effort in the so-called greedy step, in a way similar to
what was done recently via AlphaGo-Zero. I will then describe analytical
results that highlight the benefits of making this extra computational
effort.

Dates: 
Friday, May 11, 2018 - 11:00
Location: 
Inria, room A00
Speaker(s): 
Bruno Scherrer
Affiliation(s): 
Inria Nancy